Anthropomorphizing Artificial Intelligence: A Corpus Study of Mental Verbs Used with AI and ChatGPT

The language used to describe technological advancement often serves as a bridge between complex engineering and public understanding. However, as artificial intelligence (AI) becomes integrated into the fabric of daily life, the metaphors and verbs used to describe these systems are coming under intense academic scrutiny. A recent study conducted by researchers at Iowa State University, Brigham Young University, and the University of Northern Colorado has revealed that while the temptation to humanize AI is prevalent in casual conversation, professional news organizations are maintaining a surprising level of linguistic discipline. The study, titled "Anthropomorphizing Artificial Intelligence: A Corpus Study of Mental Verbs Used with AI and ChatGPT," explores the phenomenon of anthropomorphism—the attribution of human traits, emotions, or intentions to non-human entities—and how it shapes our collective perception of machine learning.

The research team, led by Jo Mackiewicz, a professor of English at Iowa State University, and Jeanine Aune, a teaching professor and director of the advanced communication program at Iowa State, analyzed massive datasets to determine how frequently mental verbs like "think," "know," "understand," and "want" are paired with AI-related terms. Their findings, published in Technical Communication Quarterly, suggest that the "blurring of the line" between human cognition and algorithmic processing is a nuanced issue, influenced heavily by editorial standards and the specific context of the technological application.

The Linguistic Mechanics of AI Perception

At the heart of the study is the use of "mental verbs." These are words traditionally reserved for sentient beings to describe internal cognitive processes. When a person says, "I understand the problem," they are describing a conscious realization. When a headline reads, "AI understands the problem," it implies a level of awareness that the technology does not possess. AI systems, including Large Language Models (LLMs) like ChatGPT, function through statistical probability and pattern recognition across vast datasets. They do not "know" facts in the human sense; they predict the next logical token in a sequence based on mathematical weights.

The researchers argue that using human-centric language creates a "false impression" of AI capabilities. Phrases such as "the AI decided" or "ChatGPT thinks" can mislead the public into believing that these systems have independent agency or moral compasses. This linguistic habit can distract from the human actors—developers, data scientists, and corporate executives—who are ultimately responsible for the output and behavior of the software.

Methodology: Analyzing 20 Billion Words

To reach their conclusions, the research team utilized the News on the Web (NOW) corpus. This dataset is one of the largest linguistic resources available, containing over 20 billion words from English-language newspapers and magazines published in 20 different countries. By using a corpus of this magnitude, the researchers were able to track real-world usage patterns across a diverse geographical and cultural landscape.

The team focused specifically on the frequency of mental verbs—such as "learns," "means," "knows," "thinks," and "needs"—when they appeared in close proximity to terms like "Artificial Intelligence," "AI," and "ChatGPT." This quantitative approach allowed the researchers to move beyond anecdotal evidence and provide a data-driven look at how the media characterizes the "mind" of the machine.

A Chronology of AI Terminology and the "ELIZA Effect"

The tendency to anthropomorphize computing is not a new phenomenon. It dates back to the mid-20th century, a period often referred to as the "AI Spring." In the 1950s, early computers were frequently described in the press as "electronic brains." This set a precedent for viewing silicon-based logic through a biological lens.

In 1966, MIT professor Joseph Weizenbaum created ELIZA, a basic natural language processing program designed to mimic a Rogerian psychotherapist. Despite the program’s simplicity, Weizenbaum was shocked to find that users attributed deep understanding and empathy to the script. This became known as the "ELIZA Effect"—the tendency of humans to read into computer responses more meaning than actually exists.

The researchers’ current study arrives at a pivotal moment in this chronology. With the public release of ChatGPT in late 2022, the "ELIZA Effect" has been magnified on a global scale. Unlike previous iterations of AI, modern LLMs produce fluid, conversational prose that makes the use of mental verbs almost instinctive for the average user.

Unexpected Findings: The Restraint of News Writers

Contrary to the expectation that news media would be saturated with humanizing language to drive clicks or simplify stories, the study found that journalists are generally cautious. The frequency of mental verbs paired with AI was significantly lower than in everyday speech.

One of the most telling data points involved the verb "needs." This word appeared most frequently with AI, totaling 661 instances in the analyzed corpus. However, upon closer inspection, the researchers found that "needs" was rarely used in an anthropomorphic sense. Instead of suggesting a human-like desire, it referred to functional requirements. Phrases like "AI needs large amounts of data" or "AI needs human oversight" treat the technology as a tool or a process rather than a sentient being. This is comparable to saying a "car needs gasoline" or a "recipe needs salt."

For ChatGPT specifically, the verb "knows" was the most common mental verb, but it appeared only 32 times within the massive dataset. This low frequency suggests that professional writers are making a conscious effort to avoid attributing "knowledge" to a chatbot, perhaps influenced by the well-documented phenomenon of "hallucinations," where AI generates confident but false information.

The Role of Editorial Standards and the AP Stylebook

The researchers attribute much of this linguistic restraint to established editorial guidelines. Organizations like the Associated Press (AP) have updated their stylebooks to address the rise of generative AI. The AP guidelines specifically advise journalists to avoid attributing human emotions or intent to AI systems.

By adhering to these standards, news writers maintain a "responsibility gap." When a headline says, "AI needs to be regulated," the use of the passive voice often implies that human regulators must take action. Jeanine Aune noted that this grammatical choice shifts the focus back to human accountability. If a writer were to say, "The AI decided to exclude certain data," it would absolve the developers of responsibility. By avoiding such phrasing, journalists preserve the understanding that AI is a product of human engineering.

The Spectrum of Anthropomorphism

One of the study’s most significant contributions is the concept that anthropomorphism is not a binary "yes or no" state, but rather a spectrum. The researchers identified instances where the language moved closer to the "human" end of the scale.

For example, stating that "AI needs to understand the real world" is more anthropomorphic than saying "AI needs more data." The former suggests a cognitive grasp of reality, ethics, and social context—traits that are currently exclusive to biological intelligence. The study found that even within a single article, a writer might use a mix of functional and anthropomorphic terms, creating a layered and sometimes contradictory image of what AI is.

"These instances showed that anthropomorphizing isn’t all-or-nothing," Aune explained. This nuance is critical for technical communicators who must balance the need for relatable language with the requirement for technical accuracy.

Implications for Public Perception and Policy

The way we talk about AI has real-world consequences for policy and public safety. If the public perceives AI as a "thinking" entity, they may be more likely to trust it with sensitive tasks, such as medical diagnoses or legal judgments, without sufficient skepticism.

Furthermore, anthropomorphic language can lead to "automation bias," where humans over-rely on automated systems because they perceive them as more objective or intelligent than themselves. By highlighting the relative scarcity of these terms in professional news, the researchers provide a benchmark for how other sectors—such as marketing, entertainment, and corporate PR—might refine their own communications to avoid misleading the public.

Reaction from the Academic and Tech Communities

While the study focused on news writing, its implications have resonated with the broader tech community. Linguists and computer scientists have long debated the "Stochastic Parrot" theory—the idea that AI is merely a complex mirror of human data without any underlying comprehension.

Reaction to the study suggests a growing consensus that "linguistic hygiene" is necessary as AI becomes more sophisticated. If we continue to use the language of the mind to describe the functions of the machine, we risk creating a society that cannot distinguish between an algorithm and a person. This has profound implications for human rights, as "personhood" is a legal status that could eventually be muddied by the very language we use to describe our tools.

Future Research and the Path Forward

The research team from Iowa State, BYU, and the University of Northern Colorado intends to expand on these findings. Future studies may look at how specific "power verbs" influence user trust. For instance, does a user feel more confident in an AI that "suggests" a solution versus one that "calculates" a solution?

As AI continues to evolve from a specialized tool into a general-purpose utility, the language will inevitably shift. The researchers emphasize that writers, educators, and developers must stay mindful of their word choices. "For writers, this nuance matters: the language we choose shapes how readers understand AI systems, their capabilities, and the humans responsible for them," Mackiewicz concluded.

The study serves as a reminder that while machines may be able to process language, the power to define the meaning of that language remains a uniquely human responsibility. By maintaining a clear distinction between "thinking" and "processing," professional communicators can ensure that the public remains informed rather than enchanted by the illusions of artificial intelligence.